Case Study — AI / ML Tooling

Smart Trainer

A bespoke deep neural network training suite, designed and delivered for Digital Scientific UK Ltd — enabling their team to build, train, and validate clinical-grade models independently.

Smart Trainer in action — training a chromosome classification model from random initialisation to clinical-grade accuracy

97%+

Classification accuracy

cross-domain validated

24

Chromosome classes

complete human karyotype

60fps

Real-time visualisation

live feature map rendering

Why 97% here is not the same as 97% elsewhere

Published chromosome classification benchmarks almost universally report accuracy on held-out data drawn from the same laboratory batch used for training — the same equipment, the same staining protocol, the same preparation conditions.

The models trained using Smart Trainer were validated against samples from an entirely separate laboratory, prepared under different conditions with zero overlap with the training data. 97%+ in this setting is a measure of real-world clinical generalisation — not in-distribution performance on a controlled benchmark.

What cross-domain means in practice

Validated on data from a different lab — different equipment, different preparation method

Zero overlap between training and validation datasets

Performs in real clinical environments, not just controlled benchmarks

Deployable across different hospital and laboratory workflows without retraining

The Challenge

Digital Scientific UK Ltd required not only a high-accuracy chromosome classification model, but the capability to build, train, and iterate on deep neural networks in-house — without relying on specialist ML engineers for every training run.

Existing ML tooling was either too low-level for clinical teams or too opaque to provide the real-time feedback needed to diagnose training issues and iterate effectively on medical imaging data.

What We Delivered

Smart Trainer is a bespoke desktop application for building and training deep neural networks, featuring real-time 60fps animated feature map visualisation, drag-and-drop dataset validation, and hardware-aware rendering that adapts dynamically to available GPU and CPU resources.

Built to enable non-technical users to run complete model training cycles independently — reducing the feedback loop from days to hours and putting clinical teams in direct control of their own AI pipeline.

Need bespoke ML tooling for your team?

We build training pipelines and tooling that put your team in control — no ML expertise required to operate them.